news

Auros asked 2,000 US shoppers to separate real products from AI ads, and published one accuracy number: 52%

The attitude figures in the October 16 study are broken out by country to the percentage point, while the real-versus-AI identification test arrives with a single average and no stimulus count.

By Daniel Ochoa · 18 October 2026

A consumer study released on October 16 put the question of AI ad detection to an actual test rather than a self-report, and the result is the least quoted number in the whole release. Respondents were shown ads and asked to decide whether each product was real or AI generated. When shown real products, an average of 52% identified them correctly. On a two-way forced choice, that sits next to chance.

Who was asked, by whom, and in what window

The research was commissioned by Auros and conducted by Talker Research. The global version covers 4,000 adults across the United States, the United Kingdom and Australia. The identification exercise is reported in the US portion, a poll of 2,000 general population Americans with internet access, administered online between August 26 and September 1, 2026. That is a seven week gap between fieldwork and publication, which matters in a category where the models shift monthly.

The attitude figures are far more granular than the test figures

On how people feel, the release is specific and segmented. 47% of Britons, 43% of Americans and 42% of Australians said they had felt duped by a product advertised with AI. More than half in each market said they do not trust products advertised with AI to be described accurately: 59% in Australia, 58% in the UK, 53% in the US. Encounter rates run 50% in the UK, 47% in Australia and 44% in the US, and respondents said they question whether an online ad shows a real or AI generated product about four times a week. Concern that AI generated advertising signals lower product quality came in at 42% in Australia, 42% in the UK and 36% in the US. Talker Research also reports that 61% believe at least one type of brand would be likely to use misleading or AI-focused marketing.

The identification exercise, which is the only part of the study that measures anything other than opinion, gets one published average and no matching accuracy figure for the AI-generated items. The number of ads shown, how they were produced, whether they were fully generated or AI-enhanced photography, and the order of presentation are not in the published write-ups. Those are the variables that decide what a 52% actually means.

Human detection keeps landing near a coin flip

This is now a pattern rather than a one-off. Tavus reported that 48% of 54 testers were fooled by its real-time video model, another result sitting on the chance line with a sample too small to separate from it. OpenAI and Meta told a US evidence-rules panel that no method reliably detects AI video. Three different organizations, three different methods, and the same conclusion arriving from opposite directions: detection by eye is not a usable control.

The commercial half of the finding is sharper. 44% of US adults in the study said they go out of their way to buy from brands that do not use AI in their ads, which means the penalty is applied on suspicion rather than on correct identification. A brand gets punished for looking AI generated whether or not it is.

Product rotation is precisely the shot type this now applies to. Here is our benchmark's product rotation prompt, run on Seedance 2.5.

Auros commissioned the study and Talker Research fielded it, so the figures carry the usual weight of vendor-sponsored research. The identification result is the part worth repeating, and it is the part that would need the full stimulus set published before anyone could score a model against it.

Sources: Auros via Business Wire, Talker Research.

Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.